AI Paper Atlas

Research map / Reinforcement learning / Offline RL

What Matters for Batch Online Reinforcement Learning in Robotics?

Perry Dong, Suvir Mirchandani, Dorsa Sadigh, Chelsea Finn

ICLR 2026 · Reinforcement learning › Offline RL · approach: Offline RL sample complexity theory

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use value-based RL (Q-function guidance) instead of IL or filtered-IL (ablation)batch online RL · 6 simulated manipulation tasksnormalized returnImproved
use filtered-IL instead of vanilla IL (ablation)batch online RL · 6 simulated manipulation tasksnormalized returnMixed
use implicit policy extraction (best-of-N by Q) instead of explicit AWR extraction (ablation)batch online RL · 6 simulated manipulation tasksnormalized return after iterationsImproved
use explicit AWR extraction (initial performance only) (ablation)batch online RL · 6 simulated manipulation tasksinitial normalized returnImproved
use diffusion policy instead of Gaussian policy (ablation)value-based batch online RL · 6 simulated manipulation tasksnormalized returnImproved
add implicit extraction to Gaussian policy (ablation)value-based batch online RL · 6 simulated manipulation tasksnormalized returnImproved
increase rollouts per iteration (ablation)value-based RL · 6 simulated tasksnormalized returnImproved
add Ornstein-Uhlenbeck temporally-correlated noise to rollouts (ablation)batch online RL recipe · 6 simulated tasksnormalized returnImproved
add temporally-correlated noise (ablation)batch online RL recipe · data scalingscaling with dataNo clear effect
use recipe instead of filtered-IL on real robotbatch online RL · Real-world Tape tasksuccess rate: 30% improvement over initial policy in 3 iterationsImproved
use steering baseline (Q-guided sampling from fixed policy)batch online RL · Real-world Tape tasksuccess rateHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy